TodayToday's edition
NVIDIA's DGX Spark 64GB was scheduled for sale on Oct. 23 through six OEM partners and keeps the GB10 chip and 273 GB/s of bandwidth, but HotHardware says the 128GB model now costs $6,995 and ties the rise to DRAM and NAND demand.
2 min readAI-assisted
NVIDIA's DGX Spark family as shown on its product page; the 64GB model is sold only through OEM partners.
Image: NVIDIA
Source · Press image from the NVIDIA product page, used for news reporting.
NVIDIA's DGX Spark 64GB was scheduled to go on sale on Friday, Oct. 23, from $4,999, sold only through Acer, ASUS, Dell, Gigabyte, HP and MSI, the company said in an Oct. 2 blog post.
Both DGX Spark configurations list 273 GB/s; the 64GB model starts at $4,999 per NVIDIA, while the $6,995 128GB price is HotHardware's report.
Illustration (AI-generated)
NVIDIA said the new configuration keeps the GB10 Grace Blackwell Superchip, DGX OS and the full NVIDIA AI software stack of the 128GB model. Its DGX Spark spec page lists the same 256-bit memory interface, 273 GB/s of memory bandwidth and ConnectX-7 network card at 200 Gbps for both versions. Voltgrid could not confirm a live partner order page showing stock on launch day, so the sale date is the one NVIDIA announced.
The price is the story. HotHardware puts the original DGX Spark's launch price at $3,999 and NVIDIA's later recommended price at $4,699, so the half-memory model now starts above both. HotHardware reports that the 128GB model has risen to $6,995 and ties the price rise to surging demand for DRAM and NAND. If that reading holds, the memory squeeze has reached the developer desk, and DRAM supply now sets how much a box can hold.
On a unified-memory machine, capacity decides the largest model that fits, while bandwidth decides how fast each token is produced. The 64GB unit halves the first and keeps the second at 273 GB/s, so it changes which models load, not how fast the memory feeds them. NVIDIA claims a single 64GB unit runs models of up to 100 billion parameters on device. HotHardware is more cautious: it says Qwen 3.8 27B and Nemotron 3.5 Lightning 30B-A3B still fit, and that the main loss is models of roughly 80 billion parameters or more.
NVIDIA's answer is clustering. Two 64GB units joined by a QSFP cable pool 128GB of memory, which NVIDIA says supports models of up to 200 billion parameters, with its Sync Cluster Assistant software handling discovery and network setup. NVIDIA also claims up to 1.7x the performance of a single system in its own Qwen 3.8 27B test; StorageReview reports that the baseline was one 128GB system. Voltgrid found no independent measurement of that figure. NVIDIA's spec page allows up to four units, with a 200 GbE switch needed for four. At the starting price, a pair of 64GB units also costs more than one 128GB unit at HotHardware's reported figure.
Cheaper rivals remain. HotHardware cites a GMKtec mini PC on AMD's Ryzen AI Max+ 395 at $3,650 with 128GB and $2,349 with 64GB, though HotHardware's testing of AMD's Ryzen AI Halo, a Ryzen AI Max+ 395 platform, found it slower than GB10 systems. NVIDIA is also bringing RTX Spark Windows PCs to market this month, as Voltgrid reported, which widens its local-AI lineup beyond DGX OS.
There is no independent measurement of a 64GB unit or of the two-unit 1.7x claim. HotHardware says NVIDIA has not stated whether storage shrinks on the 64GB systems; the spec page lists up to 4 TB. NVIDIA says its Sync Model Launcher, which downloads and starts a model such as Qwen3.8 27B on one Spark or a pair, arrives at the end of October. Partner stock and regional prices beyond the $4,999 start remain to be seen.
